GPU and Distributed Architecture Dr. Shiliang PU Hikivision - - PowerPoint PPT Presentation

Intelligent Video Analysis System Based on GPU and Distributed Architecture

• Dr. Shiliang PU

Hikivision Research Institute

High Resolution VS storage Complexity VS Accuracy Mass data VS efficiency

Mid-size City, about 22,000 cameras

316PB/year

Precious video content service under complex situation

Challenge in Video Surveillance

Surveillance video content analysis

Object detection

• Human
• Vehicle
• others

Feature

• Human

feature

• Vehicle

feature

Identification

• Human

body

• Face
• Vehicle

Surveillance video content understanding framework

Challenge in Video Surveillance

Traditional algorithm can understand simple or standard scene content

车型

Sun blade closed

Phone calling White Safe Belt Car Ford Fiesta 皖A？？66R Glass worn Male Clothes color Teenage Height ……

Traditional algorithm fails in such complex scene content, which is very common in public surveillance.

Revolution By Deep Learning in Surveillance

Deep Learning in Surveillance

Traditional algorithm Deep learning

Deep Learning in Surveillance

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

• verall passenger

channel indoor public area sunny day rainny day winter summer

Pedestrian detection Recall rate, fppi = 0.1

Traditional Deep learning

Deep Learning in Surveillance

Clothes type Riding Safe belt not fastened Phone calling backpack Hat Hanging bag Mask

Deep Learning in Surveillance

70 75 80 85 90 95 100 vehicle color brand model sun blade safe belt phone calling

Vehicle feature accuracy increased by Deep Learning

traditional algorithm deep learning

Deep Learning in Surveillance

Identity?

Deep Learning in Surveillance

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Rank1 Rank10 Rank20 Rank30 Rank40 Rank50

Face Recognition Rank in 1 million enroll dataset

Traditional Deep Learning

Deep Learning in Surveillance

Vehicle retrieval based on image comparison

Deep Learning in Surveillance

0.0000 0.1000 0.2000 0.3000 0.4000 0.5000 0.6000 0.7000 0.8000 0.9000 1.0000 Rank10 Rank25 Rank50 Rank100

Vehicle Image retrieval

Traditional Deep Learning

These are what we need! BUT……

Limitation on Deep Learning

• High computing performance

Objects detection in surveillance video require 2T Flops/sec, which needs support from high-performance computing hardware.

• High cost

Price of GPU-based server is significant higher than general server.

• High energy consumption

General server costs around 9000KWh per channel every year.

GPU solution based on distributed architecture

Tegra

Hikvision-Blade Server Base on GPUs

Advantage of Hikvision Blade Server

 System stability meets industry requirement based on low-cost chip, based on distributed-computing architecture.

1 1-1 1-2 1-3 1-1-1 1-1-2 1-1-3

Advantage of Hikvision Blade Server

16,000 14,000 14,000 300 550 8050 2,000 4,000 6,000 8,000 10,000 12,000 14,000 16,000 18,000 Blade Tesla M40*2 General Server

Performance/power ratio

Performance（Gflops） Power（W）  High performance  Low cost  Low power consumption

Advantage of Hikvision Blade Server

 Flexibility-for different product forms

Smart Server Smart IPC Smart NVR

Sensing Storage Application

Intelligent Video compressing standard

Surveillance Video

Compressing

Standard General Video Compressing Standard

Background frame IVA Bit rate equalization

Smart 264

Intelligent Video compressing standard

24 hours typical surveillance scene contrast rate at a consistent subjective quality case

• utdoor

busy free H.264 1855Kbps 1245Kbps Smart264 419Kbps 164Kbps Promotion 4.43 7.57 indoor busy free H.264 3448Kbps 1715Kbps Smart264 945Kbps 315Kbps Promotion 3.65 5.45

Intelligent Video compressing standard

H.264/H.265 Smart264

100% H.264-3830kbps Smart264-683kbps 17.8% H.2651920kbps 50%

Video structured description

Human: male female wear glasses riding backpack handbag Vehicle: driver driver’s sun visor copilot copilot’s sun visor safe belt fastened/not phone calling

Security Big Data framework

Non-structured data Structured data Cloud storage

01

High speed data bus

High speed data bus

Distributed file database Memory computing Data mining Fulltext database Police Traffic Other market

Big data manager platform

Big data service

Collecting mass data(video, image, alarm, GPS). Extracting structured data from video and images. Offering high speed service, like data searching, analyzing and statistics.

Cloud analysis

Advantages from Security Big Data

Small size Million data level Low speed Slow feature extraction Low accuracy Long time cost. >10 nins

Issues on traditional system

01 Cloud analysis handles mass-data

computing problem

02 Big data architecture handles

above billions level data

03 Spark memory computing

• ffers second degree service

04 Deep learning increases

computing accuracy

Smart traffic Police

City manag ement

……

Smart city

Statistic Alarm Analysis

Data inquiry

Security Big Data application

Security Big Data depth application

Case study – Billion-level image search engine

Image search based on image feature extraction and comparison, based on billion-level vehicle images.

Case study- Face recognition system

Lost elder found in 5 seconds.

Future  Multi-sensor increases data dimensions.  Unsupervised learning in video surveillance  Optimized neural network framework

THE END

HIKVISION Internal